Polygenic risk scores enhance prediction of body mass index increase in individuals with a first episode of psychosis

Abstract

Background: Individuals with a first episode of psychosis (FEP) show rapid weight gain during the first months of treatment, which is associated with a reduction in general physical health. Although genetics is assumed to be a significant contributor to weight gain, its exact role is unknown.

Methods: We assembled a population-based FEP cohort of 381 individuals that was split into a Training (n = 224) set and a Validation (n = 157) set to calculate the polygenic risk score (PRS) in a two-step process. In parallel, we obtained reference genome-wide association studies for body mass index (BMI) and schizophrenia (SCZ) to examine the pleiotropic landscape between the two traits. BMI PRSs were added to linear models that included sociodemographic and clinical variables to predict BMI increase (∆BMI) in the Validation set.

Results: The results confirmed considerable shared genetic susceptibility for the two traits involving 449 near-independent genomic loci. The inclusion of BMI PRSs significantly improved the prediction of ∆BMI at 12 months after the onset of antipsychotic treatment by 49.4% compared to a clinical model. In addition, we demonstrated that the PRS containing pleiotropic information between BMI and SCZ predicted ∆BMI better at 3 (12.2%) and 12 months (53.2%).

Conclusions: We prove for the first time that genetic factors play a key role in determining ∆BMI during the FEP. This finding has important clinical implications for the early identification of individuals most vulnerable to weight gain and highlights the importance of examining genetic pleiotropy in the context of medically important comorbidities for predicting future outcomes.

Keywords: BMI; first-episode psychosis; pleiotropy; polygenic risk scores; weight gain.

Figure 1.

Pleiotropic variants between SCZ and BMI. (A) Manhattan plot showing independent (r ² < 0.1) loci associated with both SCZ and BMI, as defined by conjunction false discovery rates (conj. FDR) after excluding SNPs in the MHC region. The dashed black line represents the conj. FDR threshold of 0.05. (B) Conditional Q–Q plots of nominal versus empirical (−log10) p-values (corrected for inflation) of BMI as a function of significance with SCZ, at the level of p < 10⁻¹ (red line), p < 10⁻² (yellow line), and p < 10⁻³ (purple line), respectively. The blue line indicates the standard enrichment of BMI including all SNPs, irrespective of their association with the secondary trait. The gray dashed line indicates the null distribution of p-values. (C) Pleiotropy plot for independent SNPs with conj. FDR < 0.05 (n = 486) between SCZ and BMI. The conj. FDR values and the direction of the effects (z-scores) of the minor alleles are plotted for BMI (x-axis) against SCZ (y-axis). Graph regions whose effects are consistent with a positive correlation between the two traits are shaded in yellow. (C) The ratio between discordant and concordant pleiotropic variants across different conj. FDR thresholds.

Figure 2.

∆BMI variance explained in the whole dataset. Bar plots showing the variance explained by each covariate in the Clinical models for (A) BMI₁₂, (B) ∆BMI₁₂, and (C) ∆BMI₁₂ including ∆BMI₃ in the model. AP, antipsychotic drug; CPZ, equivalent doses of chlorpromazine. *p-value < 0.05; **p-value < 0.01; ***p–value < 0.001.

Figure 3.

Clinical versus PRS models in BMI. Barplots showing the Adj. R ² in BMI by the Clinical model (CLIN) and the PRS models computed using all SNPs from BMI GWAS (PRS_BMI), pleiotropic SNPs (PRS_Pleio), and nonpleiotropic SNPs (PRS_Nonpleio) in the Validation dataset (n = 157). The barplot shows predictions of BMI₁₂, ΔBMI₃, and ΔBMI₁₂ in the x-axis. Covariates included in the Clinical model were the first 10 PC, age, sex, AP drug prescribed, chlorpromazine equivalent doses, diagnose, tobacco smoking, and cannabis use. Each PRS model was compared to the performance of the corresponding Clinical model. Asterisks represent significantly improved models compared to the Clinical models (ANOVA). *p-value < 0.05; ***p–value < 0.001.